Grf engine

[dajmcdon]

@topepo Here's a PR for edits to the framework

• Added the grf engine
• Converted the quantile outputs to a vctr type with the quantile_levels as an attribute (this will allow easier creation of downstream methods since it has a class).

library(parsnip) 
library(quantreg)
#> Loading required package: SparseM

qr_spec3 <- 
  linear_reg() %>% 
  set_engine("quantreg") %>% 
  set_mode("quantile regression", quantile_level = c(.1, .5, .9))

library(modeldata)
qr_fit3 <- 
  qr_spec3 %>% fit(price ~ sqft, data = Sacramento)
qr_fit3
#> parsnip model object
#> 
#> Call:
#> quantreg::rq(formula = price ~ sqft, tau = quantile_level, data = data)
#> 
#> Coefficients:
#>                tau= 0.1  tau= 0.5   tau= 0.9
#> (Intercept) -21815.3846 5946.2931 29181.5323
#> sqft           107.6923  136.0187   190.4032
#> 
#> Degrees of freedom: 932 total; 930 residual

predict(qr_fit3, new_data = Sacramento[1:3, ])
#> # A tibble: 3 × 1
#>             .pred_quantile
#>                 <vctrs_qn>
#> 1  [68215.385, 188358.589]
#> 2 [103861.538, 251382.041]
#> 3  [63907.692, 180742.462]

# Just 1 quantile
qr_spec1 <- linear_reg() %>% 
  set_engine("quantreg") %>% 
  set_mode("quantile regression", quantile_level = .2)

qr_fit1 <- qr_spec1 %>% fit(price ~ sqft, data = Sacramento)
qr_fit1
#> parsnip model object
#> 
#> Call:
#> quantreg::rq(formula = price ~ sqft, tau = quantile_level, data = data)
#> 
#> Coefficients:
#> (Intercept)        sqft 
#> -10794.5696    113.1899 
#> 
#> Degrees of freedom: 932 total; 930 residual
predict(qr_fit1, new_data = Sacramento[1:3, ])
#> # A tibble: 3 × 1
#>             .pred_quantile
#>                 <vctrs_qn>
#> 1   [83832.165, 83832.165]
#> 2 [121298.013, 121298.013]
#> 3     [79304.57, 79304.57]


# grf random forests engine -----------------------------------------------


library(grf)
rand_forest(mtry = integer(1), trees = integer(1), min_n = integer(1)) %>%
  set_engine("grf") %>%
  set_mode("quantile regression", quantile_level = c(.1, .5, .9)) %>%
  translate()
#> Random Forest Model Specification (quantile regression)
#> 
#> Main Arguments:
#>   mtry = integer(1)
#>   trees = integer(1)
#>   min_n = integer(1)
#> 
#> Computational engine: grf 
#> 
#> Model fit template:
#> grf::quantile_forest(x = missing_arg(), y = missing_arg(), mtry = min_cols(~integer(1), 
#>     x), num.trees = integer(1), min.node.size = min_rows(~integer(1), 
#>     x), quantiles = quantile_level, num.threads = 1L, seed = runif(1, 
#>     0, .Machine$integer.max))
#> Quantile levels: 0.1, 0.5, and 0.9.

quant_rf <- rand_forest(engine = "grf") %>%
  set_mode("quantile regression", quantile_level = c(.1, .5, .9))
out <- fit(quant_rf, formula = price ~ sqft + type, data = Sacramento)
predict(out, new_data = Sacramento[1:3, ])
#> # A tibble: 3 × 1
#>    .pred_quantile
#>        <vctrs_qn>
#> 1 [61000, 207973]
#> 2 [85000, 249862]
#> 3 [65000, 207973]


# -- standard regression with grf
reg_rf <- rand_forest(mode = "regression", engine = "grf")
out <- fit(reg_rf, formula = price ~ sqft + type, data = Sacramento)
predict(out, new_data = Sacramento[1:3, ])
#> # A tibble: 3 × 1
#>     .pred
#>     <dbl>
#> 1 124396.
#> 2 169566.
#> 3 134514.

# -- classification
class_rf <- rand_forest(mode = "classification", engine = "grf")
out <- fit(class_rf, formula = type ~ sqft + price, data = Sacramento)
predict(out, new_data = Sacramento[1:3, ], type = "class")
#> # A tibble: 3 × 1
#>   .pred_class
#>   <fct>      
#> 1 Residential
#> 2 Residential
#> 3 Residential
predict(out, new_data = Sacramento[1:3, ], type = "prob")
#> # A tibble: 3 × 3
#>   .pred_Condo .pred_Multi_Family .pred_Residential
#>         <dbl>              <dbl>             <dbl>
#> 1      0.255             0.00168             0.743
#> 2      0.0584            0.0108              0.931
#> 3      0.281             0.00163             0.718

^{Created on 2024-09-03 with reprex v2.1.1}

[topepo]

I'm looking at this and I think that I'd rather have a PR that just has the new vctrs class in it. There is a lot going on in this PR.

Also, the input being one of data.frame/matrix/vector of values really complicates the internals. Can we make it input a vector of predictions and then use lapply(), c(), or similar to create a single vec_quantiles object?

[dajmcdon]

No problem. I thought about 2 separate PRs but got lazy. I'll split these.

[github-actions]

This pull request has been automatically locked. If you believe you have found a related problem, please file a new issue (with a reprex: https://reprex.tidyverse.org) and link to this issue.